Incentivizing Reasoning from Weak Supervision

Large language models (LLMs) have demonstrated impressive performance on reasoning-intensive tasks, but enhancing their reasoning abilities typically relies on either reinforcement learning (RL) with verifiable signals or supervised fine-tuning (SFT) with high-quality long chain-of-thought (CoT) demonstrations, both of which are expensive. In this paper, we study a novel problem of incentivizing the reasoning capacity of LLMs without expensive high-quality demonstrations and reinforcement learning. We investigate whether the reasoning capabilities of LLMs can be effectively incentivized via supervision from significantly weaker models. We further analyze when and why such weak supervision succeeds in eliciting reasoning abilities in stronger models. Our findings show that supervision from significantly weaker reasoners can substantially improve student reasoning performance, recovering close to 94% of the gains of expensive RL at a fraction of the cost. Experiments across diverse benchmarks and model architectures demonstrate that weak reasoners can effectively incentivize reasoning in stronger student models, consistently improving performance across a wide range of reasoning tasks. Our results suggest that this simple weak-to-strong paradigm is a promising and generalizable alternative to costly methods for incentivizing strong reasoning capabilities at inference-time in LLMs. The code is publicly available at https://github.com/yuanyige/W2SR.

Trinity-RFT: A General-Purpose and Unified Framework for Reinforcement Fine-Tuning of Large Language Models

Trinity-RFT is a general-purpose, flexible and scalable framework designed for reinforcement fine-tuning (RFT) of large language models. It is built with a decoupled design, consisting of (1) an RFT-core that unifies and generalizes synchronous/asynchronous, on-policy/off-policy, and online/offline modes of RFT, (2) seamless integration for agent-environment interaction with high efficiency and robustness, and (3) systematic data pipelines optimized for RFT. Trinity-RFT can be easily adapted for diverse application scenarios, and serves as a unified platform for exploring advanced reinforcement learning paradigms. This technical report outlines the vision, features, design and implementations of Trinity-RFT, accompanied by extensive examples demonstrating the utility and user-friendliness of the proposed framework.

Enhancing Latent Computation in Transformers with Latent Tokens

Augmenting large language models (LLMs) with auxiliary tokens has emerged as a promising strategy for enhancing model performance. In this work, we introduce a lightweight method termed latent tokens; these are dummy tokens that may be non-interpretable in natural language but steer the autoregressive decoding process of a Transformer-based LLM via the attention mechanism. The proposed latent tokens can be seamlessly integrated with a pre-trained Transformer, trained in a parameter-efficient manner, and applied flexibly at inference time, while adding minimal complexity overhead to the existing infrastructure of standard Transformers. We propose several hypotheses about the underlying mechanisms of latent tokens and design synthetic tasks accordingly to verify them. Numerical results confirm that the proposed method noticeably outperforms the baselines, particularly in the out-of-distribution generalization scenarios, highlighting its potential in improving the adaptability of LLMs.

Tree-based Models for Vertical Federated Learning: A Survey

Tree-based models have achieved great success in a wide range of real-world applications due to their effectiveness, robustness, and interpretability, which inspired people to apply them in vertical federated learning (VFL) scenarios in recent years. In this paper, we conduct a comprehensive study to give an overall picture of applying tree-based models in VFL, from the perspective of their communication and computation protocols. We categorize tree-based models in VFL into two types, i.e., feature-gathering models and label-scattering models, and provide a detailed discussion regarding their characteristics, advantages, privacy protection mechanisms, and applications. This study also focuses on the implementation of tree-based models in VFL, summarizing several design principles for better satisfying various requirements from both academic research and industrial deployment. We conduct a series of experiments to provide empirical observations on the differences and advances of different types of tree-based models.

RePO: ReLU-based Preference Optimization

Aligning large language models (LLMs) with human preferences is critical for real-world deployment, yet existing methods like RLHF face computational and stability challenges. While DPO establishes an offline paradigm with single hyperparameter $\beta$, subsequent methods like SimPO reintroduce complexity through dual parameters ($\beta$, $\gamma$). We propose {ReLU-based Preference Optimization (RePO)}, a streamlined algorithm that eliminates $\beta$ via two advances: (1) retaining SimPO's reference-free margins but removing $\beta$ through gradient analysis, and (2) adopting a ReLU-based max-margin loss that naturally filters trivial pairs. Theoretically, RePO is characterized as SimPO's limiting case ($\beta \to \infty$), where the logistic weighting collapses to binary thresholding, forming a convex envelope of the 0-1 loss. Empirical results on AlpacaEval 2 and Arena-Hard show that RePO outperforms DPO and SimPO across multiple base models, requiring only one hyperparameter to tune.

ToolCoder: A Systematic Code-Empowered Tool Learning Framework for Large Language Models

Tool learning has emerged as a crucial capability for large language models (LLMs) to solve complex real-world tasks through interaction with external tools. Existing approaches face significant challenges, including reliance on hand-crafted prompts, difficulty in multi-step planning, and lack of precise error diagnosis and reflection mechanisms. We propose ToolCoder, a novel framework that reformulates tool learning as a code generation task. Inspired by software engineering principles, ToolCoder transforms natural language queries into structured Python function scaffold and systematically breaks down tasks with descriptive comments, enabling LLMs to leverage coding paradigms for complex reasoning and planning. It then generates and executes function implementations to obtain final responses. Additionally, ToolCoder stores successfully executed functions in a repository to promote code reuse, while leveraging error traceback mechanisms for systematic debugging, optimizing both execution efficiency and robustness. Experiments demonstrate that ToolCoder achieves superior performance in task completion accuracy and execution reliability compared to existing approaches, establishing the effectiveness of code-centric approaches in tool learning.

KIMAs: A Configurable Knowledge Integrated Multi-Agent System

Knowledge-intensive conversations supported by large language models (LLMs) have become one of the most popular and helpful applications that can assist people in different aspects. Many current knowledge-intensive applications are centered on retrieval-augmented generation (RAG) techniques. While many open-source RAG frameworks facilitate the development of RAG-based applications, they often fall short in handling practical scenarios complicated by heterogeneous data in topics and formats, conversational context management, and the requirement of low-latency response times. This technical report presents a configurable knowledge integrated multi-agent system, KIMAs, to address these challenges. KIMAs features a flexible and configurable system for integrating diverse knowledge sources with 1) context management and query rewrite mechanisms to improve retrieval accuracy and multi-turn conversational coherency, 2) efficient knowledge routing and retrieval, 3) simple but effective filter and reference generation mechanisms, and 4) optimized parallelizable multi-agent pipeline execution. Our work provides a scalable framework for advancing the deployment of LLMs in real-world settings. To show how KIMAs can help developers build knowledge-intensive applications with different scales and emphases, we demonstrate how we configure the system to three applications already running in practice with reliable performance.

Talk to Right Specialists: Routing and Planning in Multi-agent System for Question Answering

Leveraging large language models (LLMs), an agent can utilize retrieval-augmented generation (RAG) techniques to integrate external knowledge and increase the reliability of its responses. Current RAG-based agents integrate single, domain-specific knowledge sources, limiting their ability and leading to hallucinated or inaccurate responses when addressing cross-domain queries. Integrating multiple knowledge bases into a unified RAG-based agent raises significant challenges, including increased retrieval overhead and data sovereignty when sensitive data is involved. In this work, we propose RopMura, a novel multi-agent system that addresses these limitations by incorporating highly efficient routing and planning mechanisms. RopMura features two key components: a router that intelligently selects the most relevant agents based on knowledge boundaries and a planner that decomposes complex multi-hop queries into manageable steps, allowing for coordinating cross-domain responses. Experimental results demonstrate that RopMura effectively handles both single-hop and multi-hop queries, with the routing mechanism enabling precise answers for single-hop queries and the combined routing and planning mechanisms achieving accurate, multi-step resolutions for complex queries.

HumanVBench: Exploring Human-Centric Video Understanding Capabilities of MLLMs with Synthetic Benchmark Data

In the domain of Multimodal Large Language Models (MLLMs), achieving human-centric video understanding remains a formidable challenge. Existing benchmarks primarily emphasize object and action recognition, often neglecting the intricate nuances of human emotions, behaviors, and speech visual alignment within video content. We present HumanVBench, an innovative benchmark meticulously crafted to bridge these gaps in the evaluation of video MLLMs. HumanVBench comprises 17 carefully designed tasks that explore two primary dimensions: inner emotion and outer manifestations, spanning static and dynamic, basic and complex, as well as single-modal and cross-modal aspects. With two advanced automated pipelines for video annotation and distractor-included QA generation, HumanVBench utilizes diverse state-of-the-art (SOTA) techniques to streamline benchmark data synthesis and quality assessment, minimizing human annotation dependency tailored to human-centric multimodal attributes. A comprehensive evaluation across 16 SOTA video MLLMs reveals notable limitations in current performance, especially in cross-modal and temporal alignment, underscoring the necessity for further refinement toward achieving more human-like understanding. HumanVBench is open-sourced to facilitate future advancements and real-world applications in video MLLMs.

A Simple and Provable Scaling Law for the Test-Time Compute of Large Language Models

We propose a general two-stage algorithm that enjoys a provable scaling law for the test-time compute of large language models (LLMs). Given an input problem, the proposed algorithm first generates $N$ candidate solutions, and then chooses the best one via a multiple-round knockout tournament where each pair of candidates are compared for $K$ times and only the winners move on to the next round. In a minimalistic implementation, both stages can be executed with a black-box LLM alone and nothing else (e.g., no external verifier or reward model), and a total of $N \times (K + 1)$ highly parallelizable LLM calls are needed for solving an input problem. Assuming that a generated candidate solution is correct with probability $p_{\text{gen}} > 0$ and a comparison between a pair of correct and incorrect solutions identifies the right winner with probability $p_{\text{comp}} > 0.5$ (i.e., better than a random guess), we prove theoretically that the failure probability of the proposed algorithm decays to zero exponentially with respect to $N$ and $K$: $$\mathbb{P}(\text{final output is incorrect}) \le (1 - p_{\text{gen}})^N + \lceil \log_2 N \rceil e^{-2 K (p_{\text{comp}} - 0.5)^2}.$$ Our empirical results with the challenging MMLU-Pro benchmark validate the technical assumptions, as well as the efficacy of the proposed algorithm and the gains from scaling up its test-time compute.